International Journal of Computer Applications (0975 8887)

Volume 9 No.7, November 2010

8
Applying Data Mining to Customer Churn Prediction in an
Internet Service Provider

Afaq Alam Khan
BGSB University
J&K, India
Sanjay Jamwal
BGSB University
J&K, India
M.M.Sepehri
Tarbiat Modares University,
Tehran, Iran.

ABSTRACT
A business incurs much higher charges when attempting to win
new customers than to retain existing ones. As a result, much
research has been invested into new ways of identifying those
customers who have a high risk of churning. However, customer
retention efforts have also been costing organizations large
amounts of resources. Same is the situation in ISP industry in
I.R.Iran. Exploiting the use of demographic, billing and usage
data, this study tends to identify the best churn predictors on the
one hand and evaluates the accuracy of different data mining
techniques on the other. Clustering users as per their usage
features and incorporating that cluster membership information
in classification models is another aspect which has been
addressed in this study.
General Terms
Data Mining & Churn
Keywords
ISP, Churn, Data mining, Decision tree, Regression, Neural
network.
1. INTRODUCTION
Internet users worldwide are increasing at impressive rate.
According to the statistics published by Miniwatts Marketing
Group there are 1,043,104,885 internet users in the world with
8.4 million unique websites. Internet Service providers (ISP)
the companies that sell internet access in various packages, are
trying hard to add more and more customers to their system as
this helps them to keep the price of service low and thus results
in attracting further more customers [9]. Offering a variety of
services to the customers is yet another way of ISPs to increase
customer base. Customers on the other side come face to face
with a decision of choosing an ISP. They become more
demanding and tend to switch from previous service provider to
another service provider. This gives birth to the notion of churn.
Churn or customer attrition is defined as the annual turnover of
the market base [23]. Keeping in view that cost of obtaining new
customer is five times higher than maintaining an existing
customer [8] and ISPs spent a huge amount on advertisement to
gain a customer, ISPs do not afford to lose their existing
customers and need to fight customer churn.
There are two basic approaches to fight customer churn.
Untargeted approaches rely on superior product and mass
advertising to increase brand loyalty and retain customers.
Targeted approaches rely on identifying customers who are likely
to churn, and then either provide them with a direct incentive or
customize a service plan to stay. There are two types of targeted
approaches: reactive and proactive. Adopting a reactive approach,
a company waits until customers contact the company to cancel
their (service) relationship. The company then offers the
customer an incentive, for example a rebate, to stay. Adopting a
proactive approach, a company tries to identify customers who
are likely to churn at some later date in advance. The company
then targets these customers with special programmes or
incentives to keep the customers from churning. Targeted
proactive programs have potential advantages of having lower
incentive costs and because customers are not trained to
negotiate for better deals under the threat of churning. However
these systems can be very wasteful if churn predictions are
inaccurate, because then companies are wasting incentive money
on customers who would have stayed anyway. That is why the
customer churn prediction process need to be as accurate as
possible [6].
2. REVIEW OF LITERATURE
Customer attrition is an important issue for any company and is
easiest to define in subscription based businesses, and partly for
that reason, churn modeling is most popular in theses businesses
[5]. Long-distance companies, Mobile phone service providers,
Insurance companies, Cable companies (Pay-TV) [6], financial
service companies, Internet service providers, newspapers,
magazines, and some retailers all share subscription model
where customers have a formal, contractual relationship which
must be explicitly ended.
2-1. Customer Churn in Mobile service
providers:
As deregulation, new technologies, and new competition have
opened up telecommunication industry, the telecommunication
service market has become more competitive than ever. And in
this strongly competitive and broadly liberalized mobile
telecommunication industry, customer churn has turned into very
serious issue. Many subscribers frequently churn from one
provider to another in search of better rates/service or for the
benefits of signing up with new carrier (e.g., such as receiving
the latest cellular phone). Low signup fees particularly for
prepaid mobile services also encourage customers to churn.
Wireless local number portability (WLNP) or the ability to
change mobile carriers and keep mobile phone number also
poses a big challenge to the already reeling telecommunication
company.
It is estimated that the average churn rate for the
mobile telecommunications is 2.2% per month. i.e., about 27%
of given carriers subscriber are lost each year [31], making it
International Journal of Computer Applications (0975 8887)
Volume 9 No.7, November 2010
9
essential to develop an effective churn reduction method. The
cost of acquisition of new mobile service subscriber is estimated
to be from $300 to $ 600. However, the cost of retaining an
existing subscriber is generally much lower than that [30].
Wei, et al [30] used the call pattern changes and contractual data
and developed a data mining based churn prediction technique
that identify potential churners at the contract level. In their
study they used the decision tree approach as the basis for the
development of their technique for mobile telecommunication
company in Taiwan. In an another attempt to identify the
determinants of subscriber churn and customer loyalty in Korean
mobile telephony market; Kim et al [17] used binomial logit
model based on the survey of 973 mobile users .
2-2. Customer churn in banking and
insurance:
Over the past two decades, the financial markets have become
more competitive due to the mature nature of the sector on the
one hand and deregulation on the other, resulting in diminishing
profit margins and blurring distinctions between banks, insurers
and brokerage firms. Hence, now a day a small number of large
institutions offering a wide set of services dominate the financial-
service industry. These developments stimulated bank and
Insurance companies to implement customer relationship
management (CRM). Under this intensive competitive pressure,
companies realize the importance of retaining, their current
customers. The substantive relevance of attrition modeling comes
from the fact that a bank is able to increase its profits by 85%
due to a 5% improvement in the retention rate [26].
Bart et al [32] studied the defectors of saving and
investment (SI) customers of a large Belgian financial service
provider. Their study involves the use of Kplan-Meier estimator
to gain insight into the timing of the SI churn event. A multi
dimensional probit model and a proportional hazard model are
performed to find the most convenient products to cross sell in
terms of customer preferences and the likelihood to lower the
customer defection proness respectively.
In order to establish a system that can provide warning signs to
network banking services soon before they are losing valuable
customers, Ding-An Chaing et. al. [35] took association rules in
use and proposed an efficient algorithm called goal- oriented
sequential pattern, which can find out behavior patterns of
loosing customers or clues before they stop using some products.
2-3. Customer churn in ISP:
Internet popularity is growing at impressive rate. Sooner or later,
every customer comes face to face with a decision of choosing an
ISP. ISPs endure a five times higher churn rate cumulating to
10% monthly [23]. It is also said that nearly half of all Internet
subscribers leave their provider every year [2]. IP network
metering, which has lately pick up steam, is the key enabler of
getting the fundamental information of churn prediction. Shen-
Tun Li et. al. [33] applied the well known data mining
methodology CRISP-DM to investigate network usage behaviors
of the ISP subscribers in Taiwan. They used Attribute-oriented
Induction (AOI) method for discovering characteristics and
discrimination knowledge of ISP customers from the ISP traffic
data. An empirically tested model developed by Jyh-Shen Chiou
[34] examines the antecedents of consumer loyalty toward ISPs.
In choosing an ISP, pricing was and is a main differentiating
factor in mind of customers; however, service quality has
emerged as a major concern among users lately. Management of
ISP has discovered that service quality is important not only for
attracting new customers, but more importantly for retaining
existing customers who may otherwise be lured away by lower
fees. [20] developed a probability model for ISP subscriber
churn. Their model relates the probability of subscribers churn to
various service attributes and subscriber characteristics.
3. ISP Industry in Iran
In recent years, government of I.R.Iran has taken extensive and
vigorous actions to increase access to Internet. One of the
objectives of the next five year development plan is to raise
internet penetration rate from current 9.74% to 30% state wide.
To meet the above mentioned objective, four layer architecture
shown in figure 1, has been deployed by government to provide
different internet services to different kinds of users.



(Figure 1: Four layer Architecture of internet services in I.R.Iran)

ICP(Internet Connection Providers): The first tier services are
provided by ICPs. The role of ICP is to buy Internet connections
in large volumes from international vendors. They distribute the
capacity in large chunks to ISPs and volume consumers. The
second tier of Internet Services includes three groups of
providers which are as follows:
PAP(Public Access Provider): The group via DSL links provides
High speed access for their customers.
IDC(Internet Data Centre): Provides hosting services with
creating secure storage spaces and offering software services on
the internet network.
ISP(Internet service Provider): provide opportunity for
inexpensive access to internet via telephone lines (WSIS, 2005).

Table 1: Number of issued permits
VOIP IDC PAP ICP ISP
59 3 13 31 658

International Journal of Computer Applications (0975 8887)
Volume 9 No.7, November 2010
10
Most of Internet users in Iran use dial-up internet connection.
Users connect to their service provider through their phone line
and need to authenticate before they can use any service. Internet
service providers on the other hand distribute the authentication
cards in the market. Each ISP has different types of cards with
different prices. As there are more than 50 ISPs alone in Tehran,
the capital city of Iran (www.irantelecom.ir), these ISPs have
filled the market with different types of authentication cards and
offer various discounts and incentives every now and then, and
thus encourage the users to switch from one service provider to
another. Taking in consideration the large number of ISPs, and
more or less the saturation in this industry, ISPs need to shift
their focus from customer accusation to the retention of
customers, in particular the churn prediction and for this ISPs
need to know their customers and their purchase history and
usage pattern. But unfortunately most of the ISPs in Iran have no
profile of their dial-up users and does not log their purchases and
usage. In this way every time user buys a new authentication card
and connects to the ISP, he is treated as a new customer.
Sepanta Co. is one of the ISPs who operate in Tehran
and has solved the above mentioned problem by encouraging its
users to register all the authentication cards they purchase in
their user area so that they can participate in the lucky draw
every night at 00:00 hours. Sepanta Co. thus has the profile of
its dial-up users including their purchases at different times and
above all log the usage of each customer. Like all other service
providers Sepanta Co. also suffers from the customer churn.
Average monthly churn of 22% was estimated for this ISP.
Applying KDD methodology, this study exploit the use of data
provided by Sepanta Co. and presents a churn prediction model
to meet the objectives mentioned below:
a) Detection of as many churners as possible.
b) To determine the best predictors among demography, billing
and usage features.
c) To compare different techniques of classification.
Although the Said ISP offers different services including dial-up,
DSL, VOIP etc, but we used dial-up users only in this study.
First of all Dial-up being the prepaid service is ideal for
performing churn prediction, secondly there is a fierce
competition for this service and finally no usage data was
available for VOIP customers.
4. METHODOLOGY
KDD is defined as the nontrivial process of identifying valid,
noval, potentially useful and ultimately understandable patterns
in data. Based on the kind of the knowledge which can be
discovered in databases, KDD techniques can be broadly
classified in to several categories, including classification,
clustering, Association, estimation, etc. In view of the fact that
we deal with the discrete valued target variable, and our ultimate
goal is to declare each customer potentially churner or potentially
nonchurner, so the KDD function of our problem is defined to be
a classification problem [10]. Following a typical KDD
framework [13], where data mining is the core in the overall
process, the experiment went through all the steps mentioned
below.

4.1 Data works
Sepanta Co. put three of its main tables at our disposal.
tblSubscriber, tblCards and tblUsage are the three tables
containing 65000, 400000 and 15000000 records respectively. As
it is completely optional for users to register their cards and
participate in lucky draw, so large number of rows were found in
tblUsage whose related information was not available in other
two tables and were of no use for our work. tblUsage table
contain the data from 21/3/2005 to 21/5/2006 while there were
subscribers who belong to the date before 21/3/2005 also present
in the tblSubscriber table, which were again of no use for our
analysis because their usage data was not present. Similarly
declaring churn status of the subscribers who registered after
21/11/2005 was not possible. So as a first step, all those records
from the three tables were deleted which protect us to define
relationship between these tables. In this way tblSubscriber,
tblCards and tblUsage were reduced to 9148, 61993 and
2093120 records respectively.
Customers whose registration date lies with in a same
month of Persian calendar were selected from tblSubscriber table
and their purchase were searched in tblCards table after a certain
date (i,e after six months from the date of their registration). If
customer still had a purchase he was declared as nonchurner else
churner.

Figure 2: Declare churn status
As shown in figure 2 all those customers whose registration date
lies between 21/3/2005 and 21/4/2005 were selected from
tblSubscriber table and it was searched in tblCards table whether
they have any purchases after 21/09/2005. Those who still had a
purchase were declared nonchurners and rest of them were
churners. Same was done with the subscribers who registered in
the next seven months. Only 8 months from the tblSubscribers
was taken in use because it was not possible to set the churn
status for the rest of the subscribers. That is why the subscribers
who registered after 21/11/2005 were deleted from the system
including their cards and usage data.

4.2. Feature extraction
Features which have been used for churn prediction in most of
the related works can be categorised in to demographic, billing
International Journal of Computer Applications (0975 8887)
Volume 9 No.7, November 2010
11
and usage features [10][20][30]. Besides the straight forward
demographic and billing features extracted from the given
attributes, some usage features were constructed which are worth
to explain here. Whole life time of each subscriber was divided
into subperiods (15 days each) and total number of minutes
(MOU) a subscriber connected to the system during each
subperiod was calculated. Only 12 latest subperiods were taken
into consideration. It will not be inappropriate to mention here
that more recent usage behavior would be more useful for churn
prediction [30]. Frequency of use (FOU) i.e. the number of times
a subscriber connected to the system during each subperiod and
the Volume of data (VOD) transferred during each subperiod
were also calculated. Instead of adding these 36 usage attributes
to the dataset, k-means clustering technique was taken in use and
all the subscribers were clustered three times as per MOU, FOU
and VOD. In this way we add three features to our dataset which
represent the usage behaviour of the subscriber. Incorporating the
cluster membership information to the classification model has
shown better performance in the previous work [15][26]. Number
of clusters chosen for all the three times was 7 as it showed the
highest chi-square value. Portion of use (POU) is another feature
related to the different portions of the day. Usage timing for
different users is different. Some are day users and some others
connect to internet at nights. Total usage of a subscriber was
divided into six different portions (Table 2) to capture the usage
behavior of subscribers.

Table 2: Six Usage Portions of the day
Day Portion Time
From To
Evening 18:00 21:59
Latenight 22:00 01:59
Midnight 02:00 05:59
Morning 06:00 09:59
Midday 10:00 13:59
Afternoon 14:00 17:59

All the subscribers were once again clustered as per six
attributes of POU and the cluster membership information was
once again incorporated to the classification model.
4.3. Modelling
Microsoft Business intelligence development studio was used to
perform the task of classification. An object called Percentage
Sampling shown in figure 3 let us divide the dataset into
training and test sample.


Figure 3: Data flow diagram in MS Business Intelligence
Development Studio

As was recommended in the previous work, we assigned 70% of
data to the training set and the remaining 30% to the test set.
Taking use of a training set three models were built using
Microsoft Decision Trees, Microsoft Neural Network and
Microsoft logistic regression. Later the test set was used to
evaluate the accuracy of each model. Results returned by each
comes next:
4.2.1 Microsoft Decision Trees which is a hybrid decision
tree algorithm developed by Microsoft research and supports
classification and regression tasks build the tree out of training
data set. Results returned when the model was run on test set are
shown by the confusion matrix in table 3.

Table 3: Decision tree performance on test data set
Predicted
Observed
Accuracy
Churner NonChurner
Churner 1551
89.70%
151
15.79%
87.74%
Non
Churner
178
10.29%
805
84.20%
Total 1729 956

The decision tree algorithm used for the prediction of churners
returned 89.70 % true positives and 84.20 % true negatives
which accounts for the overall accuracy of 87.74 %. 178 out of
1729 cases were predicted as non churners while they were
churners. There were also 151 cases predicted as churners who
were actually non churners.
4.2.2 Microsoft logistic regression algorithm recorded the
results shown in table 4. This algorithm showed the overall
accuracy of 89.01% with 91.03 % true positives and 85.35% true
negatives
International Journal of Computer Applications (0975 8887)
Volume 9 No.7, November 2010
12
Table 4: Logistic regression performance on test data
Predicted
Observed
Accuracy
Churner NonChurner
Churner 1574
91.03 %
140
14.6 %
89.01 %
Non
Churner
155
8.96 %
816
85.35 %
Total 1729 956

4.2.3 Microsoft Neural Network algorithm is a feed
forward network and uses back propagation to adjust the
weights. Weighted sum approach is used for combination
function. While tanh function is used as activation function for
hidden nodes, sigmoid function is used for output nodes.
Microsoft neural network does not allow more than one hidden
layer and as stated below:
Number of hidden neurons = c * sqrt (m*n)
where n is the number of input neurons, m is the number of
output neurons, and c is a constant. Number of nodes in a
hidden layer can be changed by giving different values to c. This
constant is supplied to the algorithm by giving value to hidden-
node ratio. Different values were set to this parameter but the
best performance was shown when this constant was set to 4.
Table 5 shows the performance of the model. 90.74% cases were
predicted as churners who were actually churners and 86.08%
cases were predicted as non churners and were actually non
churners, which accounts for the overall accuracy of 89.08%
Table 5: Neural network performance on test data set
Predicted
Observed
Accuracy
Churner NonChurner
Churner 1569
90.74 %
133
13.91 %
89.08 %
Non
Churner
160
9.25 %
823
86.08 %
Total 1729 956

4.4 Accessing Models
The results of three algorithms; Decision trees, Logistic
regression and Neural Network is shown briefly in table 6. To
meet the first and third objectives of this study we tried three
techniques to get the best results but all the three models showed
almost similar performance with logistic regression showing a
little bit higher churner hit rate of 91.03%. Overall accuracy of
neural network is slightly higher as compared to other two
techniques.
Table 6: Summery of results
Decision
tree
Logistic
regression
Neural
Network
Churner hit rate 89.70% 91.03 % 90.74 %
Non churner hit
rate
84.20% 85.35 % 86.08 %
Accuracy 87.74% 89.01 % 89.08 %
In order to compare the over all performance of the
three techniques used, it is customary to calculate other measures
of accuracy as well. These measures are defined below:
Sensitivity = True positive / positive
Specificity = True negative/ negative
Precision = true positive / (true positive + false positive)
negtive positive
negtive
y specif icit
negitive positive
positive
y Sensitivit Accuracy

Formula Figure :3.1
In the table 7 the three algorithms are compared as per the above
defined measures.
Table 7: Other measure of accuracy
Decision
tree
Logistic
regression
Neural
Network
Sensitivity .897 .910 .907
Specificity .842 .853 .860
Precision .911 .918 .921
Accuracy .877 .890 .890

Next comes the lift chart which represents graphically the
changes that a mining model causes. This chart is an ordinary
tool to compare different models when the target is a discrete
variable (such as a customer being churner or a non churner).
The standard lift chart contains a single line for each model.

International Journal of Computer Applications (0975 8887)
Volume 9 No.7, November 2010
13

Figure 4: Accessing the models
5. DISCUSSION AND CONCLUSION
First and the third objective of this study were
discussed in the above section. This section is mainly focused on
the second objective of this study. As mentioned before, all the
features used for the churn prediction were either demographic
features or billing or usage features. Purpose of our second
objective was to get an insight about the importance of these
three types of features in the churn prediction. Having a look on
the results returned by all the three techniques, we came to this
conclusion that the demographic features have the lowest affect
on the churn prediction. But it was not easy to get to the
conclusion about the billing and usage features. It seems that
these two types are equally important for predicting churn.
Figure 5 is called the dependency network and depicts the
importance of features. The number assigned to a feature shows
the importance of that feature in churn prediction. Dependency
network is the result of decision tree algorithm and preference
given to the features is a bit different from other algorithms, but
on the whole, Usage and billing features stand at the first place
and demography has very little affect on the churn prediction.
. Figure 5: Decision tree Dependency network
Cluster membership incorporation also seems to be important as
all the three algorithms have taken them in use which brought us
to the conclusion that the clusters carry the significant
information about the usage pattern. Ignoring the incorporated
clusters during model building dropped the accuracy of all the
three algorithms considerably.
6. REFERENCES
[1] (2005). General Overview of the ICT in Iran. Tunis, World
Summit on the information society (WSIS).
[2] Au, T., Li, S., and Ma, G., (2003). "Applying and Evaluating
to Predict Customer Attrition Using Data Mining
Techniques." Journal of Comparative International
management 6(1).
[3] Au, W.-H., Chan, K. C. C., and Yao, X., (2003). "A Novel
Evolutionary DataMining Algorithm With Application to
Churn Prediction." IEEE Transactions on Evolutionary
Computation 7(6): 532-545.
[4] Baesens, B., Verstraeten, G., Van den Poel, D., Egmont-
Peterson, M., Van Kenhove, P., and Vanthienen, J., (2004).
Bayesian network classifiers for identifying the slope of the
customer lifecycle of long life customers. European Journal
of Operational Research 165: 508-523.
[5] Berry, J. A., and Linoff, G. S., (2004). Data Mining
Techniques For Marketing, Sales, and Customer
Relationship Management. USA, Wiley.
[6] Burez, J., and Van den Poel, D., (2006). "CRM at a pay-TV
company: Using analytical models to reduce customer
attrition by targeted marketing for subscription services."
Expert Systems With Applications 101: 512-524.
[7] Chae, Y. M., Hee Ho, S., Won Cho, K., Lee, D. H., and Ha
Ji, S., (2001). Data mining approach to policy analysis in a
health insurance domain. International journal for Medical
Informatics 62: 103-111.
[8] Chiang, D.-A., Wang Y.-F, and Lee, S.,-L., (2003). "Goal-
oriented sequential pattern for network banking churn
analysis." Expert System with Applications 25: 293-302.
[9] Chiou, Jyh-Shen., (2003). The antecedents of customers
loyalty toward Internet Service provider. Information &
Management 41: 685-695
[10] Daskalaki, S., Kopanas, I., Goudara, M., and Avouris, N.,
(2002). "Data mining for decision support on customer
insolvency in telecommunication business." European
Journal of Operational research 145(10): 239-255.
[11] Gappert, C., (2002). "Customer Churn in the
Communications Industry." A KPMG LLP white paper, U.S
member of KPMG International.
[12] Hadden, J., Tiwari, A., Roy, R., and Ruta, D., (2005).
"Computer assisted customer churn management: Stat-of-
the-art and future trends." Computers & Operations
Research 10(2): 139-158.
[13] Han, J., Kamber, M., (2003). Data Mining: Concepts and
Techniques, Morgan Kaufmann.
[14] Hsieh, Nan-Chen., (2004). An integrated data mining and
behavioral scoring model for analyzing bank customers.
Expert Systems with Applications 27: 623-633.
[15] Hung, S., -Y., and Wang, H.-Y., (2004). Applying Data
Mining to Telecom Churn Management. Proceedings of
International Journal of Computer Applications (0975 8887)
Volume 9 No.7, November 2010
14
third annual conference, Pacific Asia Conference on
Information Systems, Taiwan.
[16] Hwang, H., (2004). "An LTV model and customer
segmentation based on customer value: a case study on the
wireless telecommunications industry." Expert Systems with
Applications 26: 103-120.
[17] Kim, H.-S., and Yoon, C.-H., (2004). "Determinants of
subscriber churn and customer loyalty in the Korean mobile
telephony market." Telecommunication Policy 28(6): 751-
765.
[18] Lariviere, B., and Van den Poel, D., (2004). "Investigating
the role of product features in preventing customer churn by
using survival analysis and choice modeling: The case of
financial services." Expert Systems with Applications 27(2):
277-285.
[19] Li, S.-T., Shue, L.-Y., and Lee, S.-F., (2005). "Enabling
Customer Relationship Management in ISP Services
Through Mining Usage Patterns." Expert System with
Applications 100(23): 1-12.
[20] Madden, G., Scott J. S., and Cobel-Neal, G., (1999).
"Subscriber Churn in Australian ISP Market." Information
Economics and policy 11(12): 195-207.
[21] Malhotra, R., Malhotra, D. K., (2003). Evaluating
consumer loans using neural networks. International
journal of management science 31: 83-96.
[22] McKay, H., (2005). "Development of contractual
relationship between an ISP and its customers - Is a fairer
deal in sight?" Computer Law and Security Report 21: 216-
225.
[23] Minguel, A. P. M. L., (2005). "Measuring the impact of data
mining on churn management." Internet Research:
Electronic Network Applications and Policy 11(5): 375-387.
[24] Mitra, S., Acharya, T., (2003). Data Mining: Multimedia,
Soft Computing, and Bioinformatics. USA, John Wiley &
Sons.
[25] Ng, K., Liu, H., (2001). "Customer Retention Via Data
Mining." Issues on the Application of data Mining: 569-590.
[26] Prinzie, A., Van den Poel, D., (2005). Incorporating
sequential information into traditional classification models
by using an element/position-sensitive SAM. Decision
Support Systems.
[27] Van den Poel, D., (2003). "Predicting Mail-Order Repeat
Buying: Which Variables Matter?" Tijdschrift voor
Economie and management 48(3): 371-403.
[28] Van Den Poel, D., Lariviere, B., (2003). Customer attrition
analysis for financial services using proportional hazard
models. European Journal of Operational Research 157:
196-217.
[29] Verhoef, P. C., Donkers, B., (2001). Predicting customer
potential value an application in the insurance industry.
Decision Support System 32: 189-199.
[30] Wei, C.-P., and Chiu, I.-T., (2002). "Turning
telecommunications call details to churn prediction: a data
mining approach." Expert Systems with Applications
23(10): 103-112.
[31] Yan, L., Wolniewicz, R. H., and Dodier, R., (2004).
"Predicting Customer Behavior in Telecommunications."
IEEE Intelligent Systems: 50-58.
[32] Bart Lariviere and Dirk Van Den Poel (2004), Investigating
the post-complaint period by means of survival analysis,
Volume 29 , Issue 3,667-668
[33] Shen-Tun Li, shu ching kuo , Knowledge discovery in
financial investment for forecasting and trading strategy
through wavelet-based SOM networks, Expert Systems
with applications, 34(2): 935-951
[34] Jyh-Shen Chiou (2004), The antecedents of consumers
loytality towards Internet Service Provider , Information
& Management, 41(6): 685 695
[35] Ding-An Chaing et. al. (2003), A recommender system
to avoid customer churn , Expert systems with
applications, 36(4): 8071-8075